RepProbe turns repetition count into a controlled variable and finds that generalist video language models collapse at high counts even though the count remains linearly decodable from the same final-layer state, locating the failure in the output pathway rather than the representation.
The work defines the temporal divergence vector τ_l, the layer-wise representational difference induced by reversing frame order, and finds across Qwen2.5-VL-7B, Qwen3-VL-8B, and InternVL2.5-8B that its normalized magnitude peaks at intermediate-to-late layers and then declines toward the output; question-conditioned analysis and attention knockout confirm the peak is specific to temporal reasoning and functionally critical, motivating Temporal Activation Injection (TAI), a training-free method that extracts τ_l at the peak and reinjects it into subsequent layers following the measured decay, consistently improving temporal reasoning across three VideoLLMs and four benchmarks with negligible impact on non-temporal tasks.
You’re all caught up